Agent Skills | NVIDIA Dynamo Documentation

Dynamo includes agent skills for development, documentation, deployment, operations, and performance optimization. Compatible coding agents discover the skills from the repository and select them based on your request. Describe the outcome you want; you do not normally need to name a skill explicitly.

The canonical sources live in .agents/skills/. The skills/ and .claude/skills/ directories are compatibility symlinks to the same content.

Development and Documentation Skills

Skill	Use for
`debug-session`	Start a structured Dynamo bug investigation and maintain a persistent worklog.
`dep-create`	Create or update a Dynamo Enhancement Proposal (DEP), implementation plan, or retroactive DEP.
`dep-status`	Check DEP status, list DEPs by lifecycle or area, and find related proposals.
`dep-update`	Triage a DEP, assign a person in charge, or advance its review and approval state.
`dynamo-agent-harness`	Drive persistent Claude Code, Codex, or OpenCode sessions through a Dynamo endpoint over Agent Client Protocol (ACP).
`dynamo-clone-hotpath-audit`	Audit Rust hot-path `.clone()` calls and, when requested, remove unnecessary clones.
`dynamo-docs`	Add, update, move, or remove Dynamo docs, examples, recipes, and translations.
`dynamo-frontend-benchmark`	Benchmark and profile the Dynamo frontend against mock workers.
`dynamo-kv-replay-parity`	Validate byte parity and paired performance for offline KV-aware replay changes.
`fern-components`	Select and use Fern MDX components when authoring or reviewing docs pages.
`fern-navigation`	Design or change Fern navigation, tabs, sections, slugs, versions, and redirects.
`graham-code-review`	Review Rust, systems, networking, or performance-critical changes with strict Dynamo standards.
`pr-monitor`	Check pull request continuous integration (CI) health, diagnose failed jobs, distinguish regressions from flakes, and flag unexpected skips.
`visual-review`	Build an interactive HTML code-review dashboard for a pull request, branch diff, or unified diff.

Deployment and Operations Skills

Skill	Use for
`deploy-dynamo-recipe`	Deploy an assigned DynamoGraphDeployment and verify it with an OpenAI-compatible smoke test.
`dynamo-interconnect-check`	Validate NIXL, Unified Communication X (UCX), and NVIDIA Collective Communications Library (NCCL) readiness for disaggregated serving over remote direct memory access (RDMA) or NVLink.
`dynamo-router-starter`	Start or patch round-robin, KV-aware, least-loaded, or device-aware routing and run endpoint checks.
`troubleshoot-dynamo`	Diagnose failed or unhealthy deployments, pods, model caches, workers, endpoints, or benchmark jobs.

Performance Optimization Skills

These skills form the optimization workflow. The loop also uses deploy-dynamo-recipe to deploy the confirmed baseline and each approved candidate.

Phase	Skill	Use for
Workload contract	`synthesize-user-workload`	Capture the workload requirements and immutable user-provided or confirmed baseline deployment.
Baseline authoring	`author-baseline-dgd`	Draft a baseline when no catalog recipe matches, then present it for explicit user confirmation.
Benchmark configuration	`configure-aiperf-benchmark`	Freeze the AIPerf workload, objective, load policy, and Kubernetes execution manifest.
Benchmark execution	`run-aiperf-benchmark`	Launch, monitor, debug, and collect one run-scoped AIPerf benchmark.
Results analysis	`analyze-aiperf-results`	Validate and normalize results, evaluate SLOs, and compare compatible candidates.
Proposal selection	`consult-perf-knowledge`	Select one evidence-backed optimization proposal from the performance rules and engine guidance.
Candidate creation	`create-optimization-hypothesis`	Apply only the selected change to create a challenger-ready deployment draft.
Adversarial review	`perform-adversarial-review`	Review a candidate for correctness, comparability, feasibility, attribution, and worthwhile GPU spend.

See the optimization loop for the full sequence and the evidence rules for benchmark validity requirements.

Usage Notes

Invoke by intent. Ask for the result you need, such as deploying a model, investigating a regression, reviewing a pull request, or optimizing against a Service Level Objective (SLO).
Set budgets for optimization. State GPU-hour, wall-clock, and failed-deployment limits before starting a long-running optimization loop.
Use goal mode for unattended runs. Answer the agent’s workload questions before enabling the harness’s goal or budget mode.
Protect credentials. Do not provide secrets in prompts or allow credentials into generated run artifacts.